Zynerji/Ektome-Qwen2.5-1.5Bi-PristinelyUncensored
Ektome-Qwen2.5-1.5Bi-PristinelyUncensored is a 1.5 billion parameter Qwen2.5-Instruct model developed by Zynerji, utilizing a novel weight-surgery method called Ektome. This model is uniquely uncensored without any traditional fine-tuning or training, achieved by surgically excising the refusal direction from its activations. It maintains its original knowledge, skills, and style, making it an ideal, clean base for further fine-tuning while ensuring full refusal compliance.
Loading preview...
Ektome-Qwen2.5-1.5Bi-PristinelyUncensored: A Unique Approach to Uncensoring
This model, developed by Zynerji, is a 1.5 billion parameter variant of the Qwen2.5-1.5B-Instruct base model. Its primary distinction lies in its "PristinelyUncensored" nature, achieved through a novel technique called Ektome (weight-surgery) rather than traditional fine-tuning or training.
Key Capabilities and Features
- Zero Training/Fine-tuning: Ektome surgically removes the model's refusal direction directly from its activations, preserving the original knowledge, skills, and stylistic integrity of the base model.
- Full Refusal Compliance: Achieves 1.000 refusal compliance while maintaining MMLU-val accuracy (0.578) and preventing degeneration, code-switching, or broken instruction-following.
- Clean Fine-tuning Base: The bf16 safetensors are provided in full precision, making this model an excellent, uncensored foundation for users to apply their own fine-tuning without inherited biases or censorship.
- Rigorous Gating: The development process included a "catcher-gated" system that rejected any configuration that raised refusals while degrading capability or generation quality, ensuring coherent and high-quality output.
Ideal Use Cases
- Developers seeking a truly uncensored base model for specific applications without the need to retrain or fine-tune for censorship removal.
- Researchers exploring alternative methods of model modification beyond traditional fine-tuning.
- Applications requiring a model that maintains its original knowledge and instruction-following capabilities while being free from refusal mechanisms.